The challenge in one page

Everything a newcomer needs to hold in their head: who, what, when, how it is scored, what the inputs look like. Deeper material lives in The competition and The data.

Who and why

Enveda Biosciences (Boulder, Colorado; CEO Viswa Colluru, CTO August Allen) hosts the contest with Kaggle. Enveda isolates and profiles natural products for drug discovery and trained PRISM, a foundation model, on about 1.2 billion MS/MS spectra. CASMI itself was founded in 2012 by Emma Schymanski and Steffen Neumann, modelled on CASP. Editions ran 2012–2017 and 2022; this is the first aimed at the wider machine-learning community. Prize pool $50,000; winning methods are to be open-sourced and the test set is meant to persist as a benchmark.

What you submit

A CSV with two columns, molecule_id and smiles. Each row holds up to 25 SMILES strings separated by semicolons, best guess first. Every hidden molecule appears exactly once. Missing columns, empty cells, duplicated ids or more than 25 guesses reject the whole file.

How it is scored

Mean reciprocal rank at 25. A guess is correct when its 14-character InChIKey prefix — the two-dimensional connectivity, with stereochemistry ignored — matches the answer after both are passed through RDKit's tautomer canonicaliser (RDKit 2026.03.3). The molecule scores 1 divided by the rank of the first correct guess, or 0 if none of the 25 is correct. The final score is the average over molecules.

Slide the true structure to the rank your model gave it
rank 3
0.333 of a possible 1.000 for this molecule
The final score is the mean of this number over all hidden molecules. Rank 1 vs rank 2 is worth as much as rank 2 vs never finding it.

When

DateEvent
2026-09-14Competition opens on Kaggle
2026-10-07This dashboard's first build
2026-12-07Entry and team-merger deadline
2026-12-14Final submission deadline (2 selected submissions count)

Five submissions per day. Two final submissions are selected for the private leaderboard.

What the hidden test is

About 400 molecules measured on a Bruker timsTOF — natural products, hypothesised natural products, their analogs, and synthetic molecules plausible in nature — none previously published. One to sixteen spectra per molecule (median three). The visible test.parquet is a placeholder: all 1,213 of its rows are verbatim copies of training spectra from the enveda-180 library, so any local score on it is meaningless. The charts on the right describe that placeholder, which nevertheless shares the hidden set's format and acquisition settings.

What you are given to train on

About 2.54 million spectra of about 276 thousand distinct structures, pooled from eleven libraries (GNPS, MassBank/MoNA, RIKEN, the Pluskal lab MS2 set, Enveda's own enveda-180 and enveda-np-examples, and others), 82 instrument labels, 121 adduct strings, both polarities. External data and pretrained models are in play on the public leaderboard; the exact rule text should be verified on Kaggle before relying on any of them.

Anatomy of the visible test set

1,213 spectra, 400 molecules, every one on a timsTOF. A decoy, but the same shape as the real thing.

Precursor adduct (how the molecule was charged)
[M+H]+
959
[M-H]-
193
[M+CH2O2-H]-
31
[M+Na]+
22
[M+NH4]+
4
[M+K]+
2
[M+Cl]-
2
Collision energy pattern per spectrum
60 eV
363
40 eV
362
merged 20/40/60 eV
298
20 eV
190
Spectra per molecule
1 spectra
55
2 spectra
97
3 spectra
99
4 spectra
106
5 spectra
24
6 spectra
14
7 spectra
2
8 spectra
2
9 spectra
1
Precursor m/z, 10 Da bins — note the narrow 245–460 window
240–250
12
250–260
23
260–270
26
270–280
58
280–290
98
290–300
80
300–310
81
310–320
101
320–330
143
330–340
141
340–350
179
350–360
73
360–370
39
370–380
42
380–390
37
390–400
32
400–410
13
410–420
14
420–430
9
430–440
5
440–450
5
Fragment peaks per spectrum
0–25
75
25–50
86
50–100
137
100–200
242
200–400
328
400–800
245
800–1600
86
1600–3300
14
Where fragments land: all fragment m/z pooled, 10 Da bins
40–50
4,635
50–60
3,830
60–70
8,571
70–80
12,197
80–90
16,406
90–100
22,510
100–110
21,560
110–120
21,247
120–130
26,379
130–140
25,734
140–150
22,991
150–160
20,433
160–170
22,223
170–180
18,173
180–190
15,349
190–200
15,642
200–210
14,214
210–220
11,577
220–230
10,268
230–240
9,148
240–250
7,393
250–260
7,306
260–270
6,371
270–280
6,078
280–290
5,486
290–300
4,686
300–310
4,778
310–320
4,855
320–330
4,174
330–340
4,742
340–350
4,224
350–360
1,691
360–370
1,049
370–380
845
380–390
761
390–400
667
400–410
188
410–420
180
420–430
123
430–440
20
440–450
66
450–460
0

A negative-mode example

50100150200250300350precursor 335.0972
[M-H]-, collision energy 20 eV. Negative-mode spectra are 19 % of the visible set and typically carry far fewer fragments.

visible test.parquet (decoy: verbatim enveda-180 rows) via HF mirror lonelyforever/enveda.casmi, 2026-10-07